--- title: 'Markov Gap: Extreme Values & Quantum Recovery' url: https://www.emergentmind.com/topics/markov-gap type: topic --- # Markov Gap: Extreme Values & Quantum Recovery “Markov gap” is not a single invariant across the arXiv literature. In one strand of probability theory it denotes the pair of observables attached to the two highest records of a one-dimensional Markovian random walk, namely the spatial gap between the first two maxima and the time interval separating their occurrence [1303.4607]. In quantum information and holography it denotes the nonnegative difference between reflected entropy and mutual information, \(\Delta_{A:B}=S_R(A:B)-I(A:B)\), with direct interpretations in recoverability, entanglement-wedge geometry, and multipartite entanglement [2107.00009, 2507.15262]. In the theory of Markov chains and Markov operators, by contrast, the nearby standard terminology is “spectral gap” or “singular-value gap,” which measures convergence and variance decay rather than either extreme-value crowding or reflected-entropic excess [2310.10876, 2606.01683]. ## 1. Terminological scope The expression appears in at least two technically distinct senses, while a third, adjacent literature studies gap notions for Markov dynamics under different names. | Context | Quantity | Definition | |---|---|---| | Random-walk extremes | \((G_n,L_n)\) | \(G_n=M_{1,n}-M_{2,n}\), \(L_n=n_1-n_2\) | | Quantum information / holography | \(\Delta_{A:B}\) or \(M\) | \(S_R(A:B)-I(A:B)\) | | Markov-chain convergence | spectral or singular-value gap | e.g. \(1-\lambda_2\) or \(\sigma_2(I-P)\) | In the random-walk setting, the object of interest is explicitly the pair \((G_n,L_n)\) attached to the first and second maxima of a one-dimensional Markovian random walker [1303.4607]. In the quantum-information setting, the Markov gap is the difference between reflected entropy and mutual information, and is also called the “Markov distance” in the Lifshitz-theory discussion [2307.12247]. In finite-state, nonreversible, and continuous-state Markov-process theory, the corresponding gap quantities are instead defined through eigenvalues or singular values of the transition operator or generator [2310.10876, 2606.01683]. This suggests that the term is field-dependent rather than a unique standardized invariant. ## 2. Extreme-value meaning for one-dimensional random walks For a one-dimensional Markovian random walk started at \(X_0=0\) and evolving by \[ X_i=X_{i-1}+\eta_i, \] with i.i.d. symmetric jumps whose characteristic small-\(k\) behavior is \[ 1-\hat f(k)\sim |a\,k|^\mu,\qquad 0<\mu\le 2, \] the first and second maxima after \(n\) steps are \[ M_{1,n}=\max_{0\le i\le n}X_i,\qquad M_{2,n}=\text{the next largest value}, \] and, if \(n_1,n_2\) are their occurrence times, the “Markov Gap” is defined by \[ G_n=M_{1,n}-M_{2,n},\qquad L_n=n_1-n_2. \] The central object is the joint probability density \[ P_n(g,l)=\mathrm{Prob}(G_n\in dg,\,L_n=l), \] together with its large-\(n\) limit \(p(g,l)\) [1303.4607]. A three-interval renewal decomposition, combined with survival and conditioned propagators, yields a closed-form double generating function for \(P_n(g,l)\). In the stationary limit, \[ P_n(g,l)\to p(g,l), \] and the ordinary generating function in \(l\) takes the form \[ \tilde p(g,z)=\sum_{l\ge 0} z^l p(g,l)=I_1(z,g)\,I_2(g), \] with \[ I_1(z,g)= z\int_0^\infty u(z,y)\,f(g+y)\,dy,\qquad I_2(g)= \int_0^\infty h(1,x)\,f(g+x)\,dx. \] The gap marginal is \[ p_{\rm gap}(g)=2\,I_1(1,g)\,I_2(g), \] while the time-interval marginal has generating function \[ \tilde p_{\rm time}(z)=\sum_{l\ge 0} z^l p_{\rm time}(l)=\int_0^\infty I_1(z,g)\,I_2(g)\,dg \] [1303.4607]. The significance of this construction is that it furnishes an exact description of the “crowding” near the global maximum in a correlated system. In classical extreme-value theory for independent variables, the emphasis is usually on the global maximum alone; here the geometry of the top two maxima becomes tractable despite Markovian correlations. ## 3. Asymptotic laws and universality in the random-walk setting The marginal distribution of the spatial gap has a universal power-law tail for \(0<\mu<2\): \[ p_{\rm gap}(g)\sim B_\mu\, g^{-1-\mu},\qquad g\to\infty, \] with amplitude \[ B_\mu=\frac{a^\mu\,\mu}{[\Gamma(1-\mu/2)]^2}. \] For \(\mu=2\), the tail remains non-universal and depends on the details of \(f(\eta)\) [1303.4607]. The time-interval marginal exhibits three regimes: \[ p_{\rm time}(l)\sim \begin{cases} A_I\,l^{-(1+1/\mu)}, & 1<\mu\le 2,\\[4pt] A_{II}\,(\ln l)\,l^{-2}, & \mu=1,\\[4pt] A_{III}\,l^{-2}, & 0<\mu<1. \end{cases} \qquad l\to\infty. \] The case \(0<\mu<1\) is characterized by a “freezing” of the tail exponent to \(2\). The mean interval \(\langle L_n\rangle\) diverges for all \(0<\mu\le 2\), scaling as \(n^{1-1/\mu}\) for \(\mu>1\) and as \(\sim \ln n\) for \(\mu<1\) [1303.4607]. In the joint regime \(g\gg 1\), \(l\gg 1\), with \(y=l\,g^{-\mu}\) fixed, the limiting law has the scaling form \[ p(g,l)\sim g^{-1-2\mu}\,\tilde p_\mu\!\bigl(l\,g^{-\mu}\bigr), \] where \(\tilde p_\mu(y)\) is integrable on \([0,\infty)\) and obeys \[ \tilde p_\mu(y)\sim y^{-1-1/\mu},\qquad y\to\infty. \] Monte Carlo simulations confirm the tail \(p_{\rm gap}(g)\sim g^{-1-\mu}\), the change in the exponent of \(p_{\rm time}(l)\), the logarithmic correction at \(\mu=1\), and the collapse of the joint histogram in the scaling variables \(g\) and \(l\,g^{-\mu}\) [1303.4607]. ## 4. Reflected entropy minus mutual information In quantum information, one begins with a tripartite pure state \(|\psi\rangle_{ABC}\), forms \(\rho_{AB}=\mathrm{Tr}_C[|\psi\rangle\langle\psi|]\), and canonically purifies \(\rho_{AB}\) to a pure state on \(AA^*BB^*\). The reflected entropy is \[ S_R(A:B)=S(AA^*)_{|(\rho_{AB})^{1/2}\rangle}, \] while the mutual information is \[ I(A:B)=S(A)+S(B)-S(AB). \] The bipartite Markov gap is then \[ \Delta_{A:B}\equiv S_R(A:B)-I(A:B)\ge 0 \] [2507.15262]. This quantity admits an exact conditional-mutual-information representation: \[ \Delta_{A:B}=I(A:B^*|B)=I(B:A^*|A). \] Accordingly, \(\Delta_{A:B}\) measures the obstruction to quantum Markovity. When \(\Delta_{A:B}=0\), there exists a recovery map \(\mathcal R_{B\to BB^*}\) — specifically the Petz or rotated-Petz map — that acting only on \(B\) recovers the full purification on \(ABB^*\) from \(\rho_{AB}\) with perfect fidelity. More generally, one has the recoverability bound \[ I(A:B^*|B)\ge -\sup_{t\in\mathbb R}\log F\!\bigl(\rho_{ABB^*},\,\mathcal R_{B\to BB^*}^{P,t}[\rho_{AB}]\bigr), \] so the magnitude of the Markov gap controls the best achievable recovery fidelity from \(B\) alone [2507.15262]. An equivalent presentation emphasizes the same identity in the canonical purification and frames \(\Delta_M=S_R-I\) as the fidelity cost of a specific Markov recovery problem. In that formulation, \[ \Delta_M(A:B)=I(A:B^*|B)=I(B:A^*|A), \] and the gap quantifies the failure of perfect recovery in a natural channel \(B\to BB^*\) [2107.00009]. ## 5. Holographic, field-theoretic, and gravitational realizations In AdS/CFT, reflected entropy is proposed to be dual to twice the area of the entanglement-wedge cross-section, while mutual information is expressed through RT areas. In the conventions summarized for the multipartite extension, \[ S_R(A:B)\simeq \frac{2E_W(A:B)}{4G_N},\qquad I(A:B)\simeq \frac{\mathrm{area}(\gamma_A)+\mathrm{area}(\gamma_B)-\mathrm{area}(\gamma_{AB})}{4G_N}, \] so that \[ \Delta_{A:B}=\frac{2E_W(A:B)-\mathrm{area}(\gamma_{AB})}{4G_N}. \] In time-symmetric pure AdS\(_3\), the gap is universally lower bounded by a \(\log 2\) contribution per endpoint of the entanglement-wedge cross-section; in the island literature the parallel statement is that the lower bound is \(\frac{c}{3}\log 2\) times the number of EWCS boundaries on minimal surfaces [2107.00009, 2211.06886]. With islands, explicit computations in the defect extremal surface model, JT gravity, and generic \(2d\) extremal black holes display the same counting rule. In vacuum states, a boundary formulation states that the lower bound is \(\frac{c}{3}\log 2\) times the number of gaps between two boundary regions. For disjoint single intervals this gives \(M\ge 2\,\frac{c}{3}\log 2\), for adjacent intervals \(M\ge \frac{c}{3}\log 2\), and if the entanglement wedge is disconnected one has \(M=0\) [2211.06886]. In \((1+1)\)-dimensional Lifshitz theory, the reflected entropy for two disjoint intervals has the explicit form \[ S_R(A:B)= \frac{1}{\sqrt{1-\eta}}\log \!\left[\frac{1+\sqrt{1-\eta}}{\sqrt{\eta}}\right] -\log\!\bigl[2\sqrt{\eta^{-1}-1}\bigr], \] the mutual information is \[ I(A:B)=\frac12\log\!\left[\frac{1}{1-\eta}\right], \] and therefore \[ \Delta_M(A:B)= \frac{1}{\sqrt{1-\eta}}\log \!\left[\frac{1+\sqrt{1-\eta}}{\sqrt{\eta}}\right] +\log\sqrt{\eta}-\log 2. \] The gap is finite, cutoff-independent, and universal; for \(\eta\to 1\), \(\Delta_M\to 1-\log 2\simeq 0.3069\), while for adjacent intervals the gap depends nontrivially on the ratio \(v=\ell_B/\ell_A\), unlike the constant adjacent-interval result in relativistic CFTs [2307.12247]. In non-inertial frames for free fermionic fields, the Markov gap \[ M(A:B)=S_R(A:B)-I(A:B) \] shows monotonic behavior with respect to acceleration for Bell, Werner, and GHZ states. It is zero at \(r=0\) for Bell and GHZ, positive for the Werner state at \(r=0\), and strictly positive for all \(r>0\), signaling tripartite correlations involving the anti-Bob mode induced by the Unruh effect [2306.17490]. For Haar-random tripartite states, every bound-entangled marginal \(\rho_{AB}\) must have strictly positive Markov gap. Conversely, states with weakly non-zero Markov gap almost surely have an undistillable marginal state — bound entangled or separable — for sufficiently large systems. The average gap exhibits distinct large-system regimes depending on \(n_{\max}=\max(n_A,n_B,n_C)\), including a strongly nonzero regime, a weakly nonzero regime, and a regime where the gap is zero with probability \(1\) [2504.04802]. ## 6. Multipartite extensions A multipartite generalization is built from a reflected multi-entropy. Given a \(q\)-partite pure state \(|\psi\rangle_{A_1\cdots A_q}\), one traces out \(A_q\), canonically purifies the remaining mixed state on doubled systems \(A_iA_i^*\), and defines the von Neumann reflected multi-entropy \[ S_R(A_1:\cdots:A_{q-1}) \] as the \(m,n\to 1\) limit of the corresponding Rényi construction. The \(q\)-partite Markov gap is then \[ \Delta_{(q)}(A_1:\cdots:A_{q-1}) \equiv S_R(A_1:\cdots:A_{q-1}) -\Bigl[\sum_{i=1}^{q-1}S(A_i)-S(A_1\cdots A_{q-1})\Bigr]. \] In holographic states, \(\Delta_{(q)}\ge 0\) by construction [2507.15262]. The proposed interpretation is geometric as well as operational. The gap vanishes if and only if the canonical-purification cross-section can be reassembled from the ordinary RT surfaces, so that no “multi-way” geometric obstruction remains. Under holographic assumptions one also has a hierarchy \[ \Delta_{(q)}(A_1:\cdots:A_{q-1}) \ge \Delta_{(q-1)}(A_1A_2:A_3:\cdots:A_{q-1}) \ge \cdots \ge \Delta_{(2)}, \] meaning that the \(q\)-partite gap controls all coarser recoveries. Operationally, a vanishing \(\Delta_{(q)}\) implies that the full \(q\)-purification can be recovered by successive local rotated-Petz maps on one region at a time [2507.15262]. The same work introduces the genuine reflected multi-entropy because \(\Delta_{(q)}\) can fail to vanish on states lacking genuine \(q\)-way entanglement. The prescriptions require the genuine reflected quantity to vanish on any pure state whose entanglement involves fewer than \(q\) parties, to be a linear combination of reflected multi-entropies and ordinary multi-entropies of lower arity, and, at \((m,n)=(2,2)\), to reproduce twice the Rényi-2 genuine multi-entropy. For \(q=4\), an explicit formula for \(GS_{R,(3)}^{(m,n)}(A:B:C)\) is given and shown to vanish on any state factorizing a single party and on any bipartite-entangled pure state [2507.15262]. ## 7. Relation to spectral-gap terminology in Markov-process theory In the broader theory of Markov chains and Markov operators, “gap” usually refers to convergence-rate quantities rather than to either of the two Markov-gap notions above. For a reversible finite-state chain with eigenvalues \[ 1=\lambda_1>\lambda_2\ge\cdots\ge\lambda_k\ge -1, \] the spectral gap is \[ \gamma=1-\lambda_2, \] and the mixing time satisfies \[ t_{\rm mix}(\varepsilon)\le \frac{1}{\gamma}\,\ln\frac{1}{\varepsilon\,\pi_{\min}}. \] In optimal prediction, a fixed positive spectral gap eliminates the logarithmic penalty that appears when the chain may have arbitrarily small gap, leading to the parametric-rate bound \(O(k^2/n)\) for large alphabets [2106.13947]. For nonreversible finite-state chains, the generator is \(L=I-P\), the spectral gap is defined as the second-smallest singular value \[ \gamma=\sigma_2(L), \] and the relaxation time is \(\tau=1/\gamma\). This relaxation time characterizes the convergence of empirical averages and can be substantially smaller than the total-variation mixing time. Cheeger-type bounds and path-argument lower bounds generalize to this setting [2310.10876]. For general Markov semigroups on \(L^2(\mu)\), the corresponding quantity is the singular-value gap \[ \gamma=\inf_{\substack{f\in\mathrm{Dom}(L)\cap L_0^2(\mu)\\ f\neq 0}}\frac{\|Lf\|}{\|f\|}, \] which yields uniform finite-time variance bounds, bounded invertibility of \(L_0\), solvability of the Poisson equation \(-Lf=g\), and asymptotic-variance formulas. In discrete time one takes \(L=P-I\) [2606.01683]. Related operator-theoretic work studies \(L_p\)-spectral gaps of Markov maps and criteria for spectral gaps of Markov operators. For Markov maps on noncommutative probability spaces, an \(L_2\)-spectral gap implies an \(L_p\)-spectral gap for \(1